SVM deallocation should wait for completion of operations on them

(Normal `Buffer`s do that implicitly.)

- For in-order queues, a simple solution would be holding a reference to the queue that was last used to operate on the SVM, and using that to `enqueue_release` when it goes out of scope.
- For out-of-order queues, we could hold on to a bounded number of events, waiting on the ones that get booted out.
- Should this be mandatory or optional behavior?